Add script to set up CUDA environment

[ccarels]

Description
Adds environment variables to a new script setup.sh.

The first iteration will be to source the CUDA environment variables.

Outcome
User-friendly setup of environment variables.

Timeline
As soon as possible.

Interfaces
Will this feature result in a backwards-incompatible interface change: No.

Expected semantic version number increment category (Please indicate x.y.z): y.

Notes
N/A.

[KAdamek]

I think that we want a different approach to setting of cuda, which is probably easily dealt with cmake. We want to detect which cuda version is installed and change the make file accordingly so the -gencode arch=XX, code=XX but also other things reflects what is installed in the system and what GPU is present. I'm not good with bash so correct me if I'm wrong, but that is not what the setup seems to be doing. Jan could help you more regarding what CUDA is doing when it is installed, I do not know much about that.

[ccarels]

@KAdamek Thanks for the feedback. Correct, the script does not "auto detect" which CUDA version is installed.

As far as I'm aware, a CUDA library user needs to set $PATH and $LD_LIBRARY_PATH first, before being able to do CUDA function calls in a terminal as it looks like you're suggesting. (See https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#environment-setup)

My plan is to start moving the build system to CMake, this could subsume a portion of this script once that is done.

[jan2nov]

Yes, currently the script is not detecting which generation of cuda device you have and that is by user selected in the Makefile accordingly. I think you can compile with more generations and cuda_capabilities to create appropriate library version for each one but I never investigate that. Nvidia quite change the way how it is installed and if they automatically setup the PATH and LD_LIBRARY_PATH. However best way is assuming the user need to set it (either in bash_profile, or bashrc, ...). By default the cuda toolkit is installed (if the user during installing not changed it) in /usr/local/cuda-{version} and then is created a symbolic link to /usr/local/cuda to the newest version of the toolkit. Not sure where else is written the version of the cuda installed, usually I am just using nvcc --version.

…

On 07/09/18 08:22, Cees Carels wrote: @KAdamek <https://github.com/KAdamek> Thanks for the feedback. Correct, the script does not "auto detect" which CUDA version is installed. As far as I'm aware, a CUDA library user needs to set |$PATH| and |$LD_LIBRARY_PATH| first, before being able to do CUDA function calls in a terminal as it looks like you're suggesting. (See https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#environment-setup) My plan is to start moving the build system to CMake, this could subsume a portion of this script once that is done. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#53 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFCoQihwOjPQGyRc6PBwNl9Tur9RZfT0ks5uYh7HgaJpZM4Wcg_z>.

[ccarels]

I think the point here is that the makefile does not work without PATH and LD_LIBRARY_PATH being set. Similarly, nvcc --version also does not work unless the user sets PATH and LD_LIBRARY_PATH. The setup.sh script is there to make the process more user-friendly.

By the sounds of it, we could use /usr/local/cuda/ in setup.sh instead of /usr/local/cuda-{version}. I have edited cc_53_setupscript to reflect this.